Locally adaptive template sizes for matching repeat images of Earth surface mass movements

This paper presents an algorithm for locally adaptive template sizes in normalized cross-correlation (NCC) based image matching for measuring horizontal surface displacements of mass movements. After adaptively identifying candidate templates based on the image signal-to-noise ratio (SNR), the algorithm iteratively looks for the size at which the maximum cross-correlation coefficient attains a local peak and the matching position gets fixed. The algorithm is tested on modeled (deformed) images and applied to real bi-temporal images of different Earth surface mass movements. It is evaluated in comparison with globally (image-wide) fixed template sizes ranging from 11 to 101 pixels based on the improvement in the accuracy of displacement estimation and the SNR of image reconstruction. The results show that the algorithm could reduce the error of displacement estimation by up to over 90% (in the modeled case) and improve the SNR of the matching by up to over four times compared to the globally fixed template sizes highly reducing the effects of geometric distortion and noise. The algorithm pushes terrain displacement measurement from repeat images one step forward towards full automation.